Tender Wang <[email protected]> 于2025年5月25日周日 19:17写道：
>
>
> Álvaro Herrera <[email protected]> 于2025年5月24日周六 17:11写道：
>
>> On 2025-May-21, Andres Freund wrote:
>>
>> > Hi,
>> >
>> > In [1] I added some verification to projection building, to check if the
>> > tupleslot passed to ExecBuildProjectionInfo() is compatible with the
>> target
>> > list. One query in merge.sql [2] got flagged.
>> >
>> > Trying to debug that issue, I found another problem. This leaves us
>> with:
>> >
>> > 1) w/ inheritance INSERTed rows are not returned by RETURNING. This
>> seems to
>> > just generally not work with MERGE
>>
>> Hmm, curious. One thing to observe is that the original source tuple is
>> in the child table, but the tuple inserted by MERGE ends up in the
>> parent table. I'm guessing that the code gets confused as to the
>> relation that the tuple in the returned slot comes from, and that
>> somehow makes it somehow not "see" the tuple to process for RETURNING?
>> I dunno. CC'ing Dean, who is more familiar with this code than I am.
>>
>
> In ExecMergeNotMatched(), we passed the mtstate->rootResultRelInfo to
> ExecInsert().
> In this case, the ri_projectReturning of mtstate->rootResultRelInfo is
> NULL, in ExecInsert(),
> the "if (resultRelInfo->ri_projectReturning)" branch will not run, so
> inheritance INSERTed rows are not returned by RETURNING.
>
> The mtstate->rootResultRelInfo assigned in ExecInitModifyTable() is only
> here:
> if (node->rootRelation > 0)
> {
> Assert(bms_is_member(node->rootRelation, estate->es_unpruned_relids));
> mtstate->rootResultRelInfo = makeNode(ResultRelInfo);
> ExecInitResultRelation(estate, mtstate->rootResultRelInfo,
> node->rootRelation);
> }
> The ri_projectReturning is not assigned.
>
> I try to pass resultRelInfo to ExecInsert, the inherited INSERTed rows are
> returned by RETURNING.
> But some test cases in regression failed.
>
For a partitioned table, we must pass rootResultRelInfo to ExecInsert(). I
added the check before calling ExecInsert()
If it is a partitioned table, we continue to pass rootResultRelInfo.
Otherwise, we pass resultRelInfo.
Please see the attached diff file. The patch passed all regression test
cases.
--
Thanks,
Tender Wang
diff --git a/src/backend/executor/nodeModifyTable.c
b/src/backend/executor/nodeModifyTable.c
index 2bc89bf84dc..443d4523789 100644
--- a/src/backend/executor/nodeModifyTable.c
+++ b/src/backend/executor/nodeModifyTable.c
@@ -3612,6 +3612,7 @@ ExecMergeNotMatched(ModifyTableContext *context,
ResultRelInfo *resultRelInfo,
MergeActionState *action = (MergeActionState *) lfirst(l);
CmdType commandType = action->mas_action->commandType;
TupleTableSlot *newslot;
+ char relkind;
/*
* Test condition, if any.
@@ -3635,9 +3636,15 @@ ExecMergeNotMatched(ModifyTableContext *context,
ResultRelInfo *resultRelInfo,
*/
newslot = ExecProject(action->mas_proj);
mtstate->mt_merge_action = action;
+ relkind =
mtstate->rootResultRelInfo->ri_RelationDesc->rd_rel->relkind;
- rslot = ExecInsert(context,
mtstate->rootResultRelInfo,
+ if (relkind == RELKIND_PARTITIONED_TABLE)
+ rslot = ExecInsert(context,
mtstate->rootResultRelInfo,
newslot,
canSetTag, NULL, NULL);
+ else
+ rslot = ExecInsert(context,
resultRelInfo,
+ newslot,
canSetTag, NULL, NULL);
+
mtstate->mt_merge_inserted += 1;
break;
case CMD_NOTHING:
